* Started from sample code by Juan J. Sierralta P.
*/
+#include "config.h"
#include <stdlib.h>
#include <stdio.h>
#include <string.h>
-#include <sys/time.h>
+#if HAVE_UNISTD_H
#include <unistd.h>
+#endif
#include <math.h>
#include "libavutil/cpu.h"
#include "libavutil/common.h"
#include "libavutil/lfg.h"
+#include "libavutil/time.h"
+#include "dct.h"
#include "simple_idct.h"
#include "aandcttab.h"
#include "faandct.h"
#undef printf
-void ff_mmx_idct(DCTELEM *data);
-void ff_mmxext_idct(DCTELEM *data);
-
-void odivx_idct_c(short *block);
-
// BFIN
-void ff_bfin_idct(DCTELEM *block);
-void ff_bfin_fdct(DCTELEM *block);
+void ff_bfin_idct(int16_t *block);
+void ff_bfin_fdct(int16_t *block);
// ALTIVEC
-void fdct_altivec(DCTELEM *block);
-//void idct_altivec(DCTELEM *block);?? no routine
+void ff_fdct_altivec(int16_t *block);
// ARM
-void ff_j_rev_dct_arm(DCTELEM *data);
-void ff_simple_idct_arm(DCTELEM *data);
-void ff_simple_idct_armv5te(DCTELEM *data);
-void ff_simple_idct_armv6(DCTELEM *data);
-void ff_simple_idct_neon(DCTELEM *data);
+void ff_j_rev_dct_arm(int16_t *data);
+void ff_simple_idct_arm(int16_t *data);
+void ff_simple_idct_armv5te(int16_t *data);
+void ff_simple_idct_armv6(int16_t *data);
+void ff_simple_idct_neon(int16_t *data);
-void ff_simple_idct_axp(DCTELEM *data);
+void ff_simple_idct_axp(int16_t *data);
struct algo {
const char *name;
- enum { FDCT, IDCT } is_idct;
- void (*func)(DCTELEM *block);
- void (*ref) (DCTELEM *block);
+ void (*func)(int16_t *block);
enum formattag { NO_PERM, MMX_PERM, MMX_SIMPLE_PERM, SCALE_PERM,
SSE2_PERM, PARTTRANS_PERM } format;
int mm_support;
+ int nonspec;
};
-#ifndef FAAN_POSTSCALE
-#define FAAN_SCALE SCALE_PERM
-#else
-#define FAAN_SCALE NO_PERM
-#endif
-
static int cpu_flags;
-struct algo algos[] = {
- {"REF-DBL", 0, ff_ref_fdct, ff_ref_fdct, NO_PERM},
- {"FAAN", 0, ff_faandct, ff_ref_fdct, FAAN_SCALE},
- {"FAANI", 1, ff_faanidct, ff_ref_idct, NO_PERM},
- {"IJG-AAN-INT", 0, fdct_ifast, ff_ref_fdct, SCALE_PERM},
- {"IJG-LLM-INT", 0, ff_jpeg_fdct_islow, ff_ref_fdct, NO_PERM},
- {"REF-DBL", 1, ff_ref_idct, ff_ref_idct, NO_PERM},
- {"INT", 1, j_rev_dct, ff_ref_idct, MMX_PERM},
- {"SIMPLE-C", 1, ff_simple_idct, ff_ref_idct, NO_PERM},
-
-#if HAVE_MMX
- {"MMX", 0, ff_fdct_mmx, ff_ref_fdct, NO_PERM, AV_CPU_FLAG_MMX},
-#if HAVE_MMX2
- {"MMX2", 0, ff_fdct_mmx2, ff_ref_fdct, NO_PERM, AV_CPU_FLAG_MMX2},
- {"SSE2", 0, ff_fdct_sse2, ff_ref_fdct, NO_PERM, AV_CPU_FLAG_SSE2},
+static const struct algo fdct_tab[] = {
+ { "REF-DBL", ff_ref_fdct, NO_PERM },
+ { "FAAN", ff_faandct, NO_PERM },
+ { "IJG-AAN-INT", ff_fdct_ifast, SCALE_PERM },
+ { "IJG-LLM-INT", ff_jpeg_fdct_islow_8, NO_PERM },
+
+#if HAVE_MMX_INLINE
+ { "MMX", ff_fdct_mmx, NO_PERM, AV_CPU_FLAG_MMX },
+ { "MMXEXT", ff_fdct_mmxext, NO_PERM, AV_CPU_FLAG_MMXEXT },
+ { "SSE2", ff_fdct_sse2, NO_PERM, AV_CPU_FLAG_SSE2 },
#endif
-#if CONFIG_GPL
- {"LIBMPEG2-MMX", 1, ff_mmx_idct, ff_ref_idct, MMX_PERM, AV_CPU_FLAG_MMX},
- {"LIBMPEG2-MMX2", 1, ff_mmxext_idct, ff_ref_idct, MMX_PERM, AV_CPU_FLAG_MMX2},
+#if HAVE_ALTIVEC
+ { "altivecfdct", ff_fdct_altivec, NO_PERM, AV_CPU_FLAG_ALTIVEC },
#endif
- {"SIMPLE-MMX", 1, ff_simple_idct_mmx, ff_ref_idct, MMX_SIMPLE_PERM, AV_CPU_FLAG_MMX},
- {"XVID-MMX", 1, ff_idct_xvid_mmx, ff_ref_idct, NO_PERM, AV_CPU_FLAG_MMX},
- {"XVID-MMX2", 1, ff_idct_xvid_mmx2, ff_ref_idct, NO_PERM, AV_CPU_FLAG_MMX2},
- {"XVID-SSE2", 1, ff_idct_xvid_sse2, ff_ref_idct, SSE2_PERM, AV_CPU_FLAG_SSE2},
+
+#if ARCH_BFIN
+ { "BFINfdct", ff_bfin_fdct, NO_PERM },
#endif
-#if HAVE_ALTIVEC
- {"altivecfdct", 0, fdct_altivec, ff_ref_fdct, NO_PERM, AV_CPU_FLAG_ALTIVEC},
+ { 0 }
+};
+
+static const struct algo idct_tab[] = {
+ { "FAANI", ff_faanidct, NO_PERM },
+ { "REF-DBL", ff_ref_idct, NO_PERM },
+ { "INT", ff_j_rev_dct, MMX_PERM },
+ { "SIMPLE-C", ff_simple_idct_8, NO_PERM },
+
+#if HAVE_MMX_INLINE
+ { "SIMPLE-MMX", ff_simple_idct_mmx, MMX_SIMPLE_PERM, AV_CPU_FLAG_MMX },
+ { "XVID-MMX", ff_idct_xvid_mmx, NO_PERM, AV_CPU_FLAG_MMX, 1 },
+ { "XVID-MMXEXT", ff_idct_xvid_mmxext, NO_PERM, AV_CPU_FLAG_MMXEXT, 1 },
+ { "XVID-SSE2", ff_idct_xvid_sse2, SSE2_PERM, AV_CPU_FLAG_SSE2, 1 },
#endif
#if ARCH_BFIN
- {"BFINfdct", 0, ff_bfin_fdct, ff_ref_fdct, NO_PERM},
- {"BFINidct", 1, ff_bfin_idct, ff_ref_idct, NO_PERM},
+ { "BFINidct", ff_bfin_idct, NO_PERM },
#endif
#if ARCH_ARM
- {"SIMPLE-ARM", 1, ff_simple_idct_arm, ff_ref_idct, NO_PERM },
- {"INT-ARM", 1, ff_j_rev_dct_arm, ff_ref_idct, MMX_PERM },
+ { "SIMPLE-ARM", ff_simple_idct_arm, NO_PERM },
+ { "INT-ARM", ff_j_rev_dct_arm, MMX_PERM },
+#endif
#if HAVE_ARMV5TE
- {"SIMPLE-ARMV5TE", 1, ff_simple_idct_armv5te, ff_ref_idct, NO_PERM },
+ { "SIMPLE-ARMV5TE", ff_simple_idct_armv5te,NO_PERM, AV_CPU_FLAG_ARMV5TE },
#endif
#if HAVE_ARMV6
- {"SIMPLE-ARMV6", 1, ff_simple_idct_armv6, ff_ref_idct, MMX_PERM },
+ { "SIMPLE-ARMV6", ff_simple_idct_armv6, MMX_PERM, AV_CPU_FLAG_ARMV6 },
#endif
#if HAVE_NEON
- {"SIMPLE-NEON", 1, ff_simple_idct_neon, ff_ref_idct, PARTTRANS_PERM },
+ { "SIMPLE-NEON", ff_simple_idct_neon, PARTTRANS_PERM, AV_CPU_FLAG_NEON },
#endif
-#endif /* ARCH_ARM */
#if ARCH_ALPHA
- {"SIMPLE-ALPHA", 1, ff_simple_idct_axp, ff_ref_idct, NO_PERM },
+ { "SIMPLE-ALPHA", ff_simple_idct_axp, NO_PERM },
#endif
{ 0 }
#define AANSCALE_BITS 12
-uint8_t cropTbl[256 + 2 * MAX_NEG_CROP];
-
-static int64_t gettime(void)
-{
- struct timeval tv;
- gettimeofday(&tv, NULL);
- return (int64_t)tv.tv_sec * 1000000 + tv.tv_usec;
-}
-
#define NB_ITS 20000
#define NB_ITS_SPEED 50000
}
}
-DECLARE_ALIGNED(16, static DCTELEM, block)[64];
-DECLARE_ALIGNED(8, static DCTELEM, block1)[64];
-DECLARE_ALIGNED(8, static DCTELEM, block_org)[64];
+DECLARE_ALIGNED(16, static int16_t, block)[64];
+DECLARE_ALIGNED(8, static int16_t, block1)[64];
-static inline void mmx_emms(void)
+static void init_block(int16_t block[64], int test, int is_idct, AVLFG *prng)
{
-#if HAVE_MMX
- if (cpu_flags & AV_CPU_FLAG_MMX)
- __asm__ volatile ("emms\n\t");
-#endif
+ int i, j;
+
+ memset(block, 0, 64 * sizeof(*block));
+
+ switch (test) {
+ case 0:
+ for (i = 0; i < 64; i++)
+ block[i] = (av_lfg_get(prng) % 512) - 256;
+ if (is_idct) {
+ ff_ref_fdct(block);
+ for (i = 0; i < 64; i++)
+ block[i] >>= 3;
+ }
+ break;
+ case 1:
+ j = av_lfg_get(prng) % 10 + 1;
+ for (i = 0; i < j; i++)
+ block[av_lfg_get(prng) % 64] = av_lfg_get(prng) % 512 - 256;
+ break;
+ case 2:
+ block[ 0] = av_lfg_get(prng) % 4096 - 2048;
+ block[63] = (block[0] & 1) ^ 1;
+ break;
+ }
+}
+
+static void permute(int16_t dst[64], const int16_t src[64], int perm)
+{
+ int i;
+
+ if (perm == MMX_PERM) {
+ for (i = 0; i < 64; i++)
+ dst[idct_mmx_perm[i]] = src[i];
+ } else if (perm == MMX_SIMPLE_PERM) {
+ for (i = 0; i < 64; i++)
+ dst[idct_simple_mmx_perm[i]] = src[i];
+ } else if (perm == SSE2_PERM) {
+ for (i = 0; i < 64; i++)
+ dst[(i & 0x38) | idct_sse2_row_perm[i & 7]] = src[i];
+ } else if (perm == PARTTRANS_PERM) {
+ for (i = 0; i < 64; i++)
+ dst[(i & 0x24) | ((i & 3) << 3) | ((i >> 3) & 3)] = src[i];
+ } else {
+ for (i = 0; i < 64; i++)
+ dst[i] = src[i];
+ }
}
-static void dct_error(const struct algo *dct, int test)
+static int dct_error(const struct algo *dct, int test, int is_idct, int speed)
{
+ void (*ref)(int16_t *block) = is_idct ? ff_ref_idct : ff_ref_fdct;
int it, i, scale;
int err_inf, v;
- int64_t err2, ti, ti1, it1;
+ int64_t err2, ti, ti1, it1, err_sum = 0;
int64_t sysErr[64], sysErrMax = 0;
int maxout = 0;
int blockSumErrMax = 0, blockSumErr;
AVLFG prng;
+ double omse, ome;
+ int spec_err;
av_lfg_init(&prng, 1);
for (i = 0; i < 64; i++)
sysErr[i] = 0;
for (it = 0; it < NB_ITS; it++) {
- for (i = 0; i < 64; i++)
- block1[i] = 0;
- switch (test) {
- case 0:
- for (i = 0; i < 64; i++)
- block1[i] = (av_lfg_get(&prng) % 512) - 256;
- if (dct->is_idct) {
- ff_ref_fdct(block1);
- for (i = 0; i < 64; i++)
- block1[i] >>= 3;
- }
- break;
- case 1: {
- int num = av_lfg_get(&prng) % 10 + 1;
- for (i = 0; i < num; i++)
- block1[av_lfg_get(&prng) % 64] =
- av_lfg_get(&prng) % 512 - 256;
- }
- break;
- case 2:
- block1[0] = av_lfg_get(&prng) % 4096 - 2048;
- block1[63] = (block1[0] & 1) ^ 1;
- break;
- }
-
- for (i = 0; i < 64; i++)
- block_org[i] = block1[i];
-
- if (dct->format == MMX_PERM) {
- for (i = 0; i < 64; i++)
- block[idct_mmx_perm[i]] = block1[i];
- } else if (dct->format == MMX_SIMPLE_PERM) {
- for (i = 0; i < 64; i++)
- block[idct_simple_mmx_perm[i]] = block1[i];
- } else if (dct->format == SSE2_PERM) {
- for (i = 0; i < 64; i++)
- block[(i & 0x38) | idct_sse2_row_perm[i & 7]] = block1[i];
- } else if (dct->format == PARTTRANS_PERM) {
- for (i = 0; i < 64; i++)
- block[(i & 0x24) | ((i & 3) << 3) | ((i >> 3) & 3)] = block1[i];
- } else {
- for (i = 0; i < 64; i++)
- block[i] = block1[i];
- }
+ init_block(block1, test, is_idct, &prng);
+ permute(block, block1, dct->format);
dct->func(block);
- mmx_emms();
+ emms_c();
if (dct->format == SCALE_PERM) {
for (i = 0; i < 64; i++) {
}
}
- dct->ref(block1);
+ ref(block1);
blockSumErr = 0;
for (i = 0; i < 64; i++) {
- v = abs(block[i] - block1[i]);
+ int err = block[i] - block1[i];
+ err_sum += err;
+ v = abs(err);
if (v > err_inf)
err_inf = v;
err2 += v * v;
}
printf("\n");
- printf("%s %s: err_inf=%d err2=%0.8f syserr=%0.8f maxout=%d blockSumErr=%d\n",
- dct->is_idct ? "IDCT" : "DCT", dct->name, err_inf,
- (double) err2 / NB_ITS / 64.0, (double) sysErrMax / NB_ITS,
+ omse = (double) err2 / NB_ITS / 64;
+ ome = (double) err_sum / NB_ITS / 64;
+
+ spec_err = is_idct && (err_inf > 1 || omse > 0.02 || fabs(ome) > 0.0015);
+
+ printf("%s %s: ppe=%d omse=%0.8f ome=%0.8f syserr=%0.8f maxout=%d blockSumErr=%d\n",
+ is_idct ? "IDCT" : "DCT", dct->name, err_inf,
+ omse, ome, (double) sysErrMax / NB_ITS,
maxout, blockSumErrMax);
- /* speed test */
- for (i = 0; i < 64; i++)
- block1[i] = 0;
+ if (spec_err && !dct->nonspec)
+ return 1;
- switch (test) {
- case 0:
- for (i = 0; i < 64; i++)
- block1[i] = av_lfg_get(&prng) % 512 - 256;
- if (dct->is_idct) {
- ff_ref_fdct(block1);
- for (i = 0; i < 64; i++)
- block1[i] >>= 3;
- }
- break;
- case 1:
- case 2:
- block1[0] = av_lfg_get(&prng) % 512 - 256;
- block1[1] = av_lfg_get(&prng) % 512 - 256;
- block1[2] = av_lfg_get(&prng) % 512 - 256;
- block1[3] = av_lfg_get(&prng) % 512 - 256;
- break;
- }
+ if (!speed)
+ return 0;
- if (dct->format == MMX_PERM) {
- for (i = 0; i < 64; i++)
- block[idct_mmx_perm[i]] = block1[i];
- } else if (dct->format == MMX_SIMPLE_PERM) {
- for (i = 0; i < 64; i++)
- block[idct_simple_mmx_perm[i]] = block1[i];
- } else {
- for (i = 0; i < 64; i++)
- block[i] = block1[i];
- }
+ /* speed test */
+ init_block(block, test, is_idct, &prng);
+ permute(block1, block, dct->format);
- ti = gettime();
+ ti = av_gettime();
it1 = 0;
do {
for (it = 0; it < NB_ITS_SPEED; it++) {
- for (i = 0; i < 64; i++)
- block[i] = block1[i];
+ memcpy(block, block1, sizeof(block));
dct->func(block);
}
it1 += NB_ITS_SPEED;
- ti1 = gettime() - ti;
+ ti1 = av_gettime() - ti;
} while (ti1 < 1000000);
- mmx_emms();
+ emms_c();
- printf("%s %s: %0.1f kdct/s\n", dct->is_idct ? "IDCT" : "DCT", dct->name,
+ printf("%s %s: %0.1f kdct/s\n", is_idct ? "IDCT" : "DCT", dct->name,
(double) it1 * 1000.0 / (double) ti1);
+
+ return 0;
}
DECLARE_ALIGNED(8, static uint8_t, img_dest)[64];
static void idct248_error(const char *name,
void (*idct248_put)(uint8_t *dest, int line_size,
- int16_t *block))
+ int16_t *block),
+ int speed)
{
int it, i, it1, ti, ti1, err_max, v;
AVLFG prng;
}
printf("%s %s: err_inf=%d\n", 1 ? "IDCT248" : "DCT248", name, err_max);
- ti = gettime();
+ if (!speed)
+ return;
+
+ ti = av_gettime();
it1 = 0;
do {
for (it = 0; it < NB_ITS_SPEED; it++) {
idct248_put(img_dest, 8, block);
}
it1 += NB_ITS_SPEED;
- ti1 = gettime() - ti;
+ ti1 = av_gettime() - ti;
} while (ti1 < 1000000);
- mmx_emms();
+ emms_c();
printf("%s %s: %0.1f kdct/s\n", 1 ? "IDCT248" : "DCT248", name,
(double) it1 * 1000.0 / (double) ti1);
" 1 -> test with random sparse matrixes\n"
" 2 -> do 3. test from mpeg4 std\n"
"-i test IDCT implementations\n"
- "-4 test IDCT248 implementations\n");
+ "-4 test IDCT248 implementations\n"
+ "-t speed test\n");
}
+#if !HAVE_GETOPT
+#include "compat/getopt.c"
+#endif
+
int main(int argc, char **argv)
{
int test_idct = 0, test_248_dct = 0;
int c, i;
int test = 1;
+ int speed = 0;
+ int err = 0;
cpu_flags = av_get_cpu_flags();
ff_ref_dct_init();
idct_mmx_init();
- for (i = 0; i < 256; i++)
- cropTbl[i + MAX_NEG_CROP] = i;
- for (i = 0; i < MAX_NEG_CROP; i++) {
- cropTbl[i] = 0;
- cropTbl[i + MAX_NEG_CROP + 256] = 255;
- }
-
for (;;) {
- c = getopt(argc, argv, "ih4");
+ c = getopt(argc, argv, "ih4t");
if (c == -1)
break;
switch (c) {
case '4':
test_248_dct = 1;
break;
+ case 't':
+ speed = 1;
+ break;
default:
case 'h':
help();
if (optind < argc)
test = atoi(argv[optind]);
- printf("ffmpeg DCT/IDCT test\n");
+ printf("Libav DCT/IDCT test\n");
if (test_248_dct) {
- idct248_error("SIMPLE-C", ff_simple_idct248_put);
+ idct248_error("SIMPLE-C", ff_simple_idct248_put, speed);
} else {
+ const struct algo *algos = test_idct ? idct_tab : fdct_tab;
for (i = 0; algos[i].name; i++)
- if (algos[i].is_idct == test_idct &&
- !(~cpu_flags & algos[i].mm_support)) {
- dct_error(&algos[i], test);
+ if (!(~cpu_flags & algos[i].mm_support)) {
+ err |= dct_error(&algos[i], test, test_idct, speed);
}
}
- return 0;
+
+ return err;
}
projects / ffmpeg / blobdiff